WeDLM-8B-Instruct-MLX

This is a full-precision (fp16) MLX version of tencent/WeDLM-8B-Instruct for inference on Apple Silicon.

Model Details

• Base Model: tencent/WeDLM-8B-Instruct
• Precision: fp16 (no quantization)
• Format: MLX SafeTensors
• Size: ~15.2 GB

About WeDLM

WeDLM (Window-based Efficient Diffusion Language Model) is a novel approach that combines:

• Entropy-based parallel decoding: Multiple tokens generated simultaneously based on prediction confidence
• Topological reordering: Efficient KV cache layout while preserving logical positions via RoPE
• Window-based generation: Fixed-size window processed in parallel per forward pass

Reference: arXiv:2512.22737

Usage

Installation

pip install mlx mlx-lm

Quick Start

from mlx_lm import load, generate

model, tokenizer = load("zimengxiong/WeDLM-8B-Instruct-MLX")
response = generate(model, tokenizer, prompt="What is machine learning?", max_tokens=256)
print(response)

Chat Template

from mlx_lm import load, generate

model, tokenizer = load("zimengxiong/WeDLM-8B-Instruct-MLX")

messages = [{"role": "user", "content": "Explain quantum computing in simple terms."}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=prompt, max_tokens=512)
print(response)

Model Architecture

Parameter	Value
Hidden Size	4096
Intermediate Size	12288
Num Layers	36
Num Attention Heads	32
Num KV Heads	8
Head Dim	128
Vocab Size	151936
Max Position Embeddings	16384
RoPE Theta	1000000

Related Models

Variant	HuggingFace
4-bit	zimengxiong/WeDLM-8B-Instruct-MLX-4bit
8-bit	zimengxiong/WeDLM-8B-Instruct-MLX-8bit
fp16 (this model)	zimengxiong/WeDLM-8B-Instruct-MLX

License

This model inherits the license from the base model tencent/WeDLM-8B-Instruct.